Z.ai's stealth ox-alpha model launches as GLM-5.3-Flash at cents per million tokens

A stealth preview becomes GLM-5.3-Flash

The unbranded model that appeared on OpenRouter.AI and OpenCode.AI under the label "ox-alpha" has been revealed as GLM-5.3-Flash, a new open-weights release from Z.ai. According to a dev.to post by developer simbo1905 published on 28 August 2026, users had already worked out that the free preview was a multi-modal Z.ai model; the genuine puzzle was how a giveaway could absorb its traffic. The author cites single-day preview usage of roughly 5.8 trillion prompt tokens and about 94 billion completion tokens.

His explanation, now that the model has launched, is that a Flash-class model this efficient simply does not cost much to serve.

Pricing and hosting

GLM-5.3-Flash is hosted on US hardware run by San Francisco firm OpenCode.AI. The dev.to post lists input pricing of $0.07 per million tokens and output pricing of $0.25 per million tokens. Because the model is open-weights, operators can also run it on their own infrastructure rather than depending on a single provider.

The joke test

The post's most unusual claim concerns jokes. Testing the model through his personal fork of a Rust-based, Apache-2.0-licensed coding agent, the author asked it for a joke and reports that it improvised one on the spot — and flopped. In his experience, most current models respond to that prompt with the same memorised handful of jokes, whatever harness they run in.

He reads the failure as a positive signal: a model that attempts fresh construction instead of reciting known material is doing something different under the hood, and he calls that behaviour a feature rather than a defect. It is one developer's informal experiment rather than a controlled benchmark, but it is a cheap probe for behavioural differences between otherwise similar assistants.

Overnight agents for pocket change

The practical case in the post is cost. The author describes running OpenCode.AI's coding agent across his current projects all night, including a cleanup of one large repository that was still going after twelve hours. He reports spending no more than $0.02 on any single bug-fix task and $0.47 in total for the run, and says an earlier overnight session came in cheaper than a coffee from his local Gail's Bakery.

Cross-checking is where the contrast shows. He says he had Kimi K3 review the work, and that it cost ten times as much just to look at the git diff. The post also points to a 45-minute review from Theo.gg that placed ox-alpha alongside Claude Opus and OpenAI Sol for long-running agentic tasks.

Why it matters

Cheap open-weights models change the economics of agentic coding. A twelve-hour autonomous run that costs under a dollar, as the author reports, makes always-on agents realistic for individuals and small teams in a way that typical per-token frontier pricing has not been. It also makes verification the budget line to watch: if a reviewer model costs an order of magnitude more than the model that did the work, the review step becomes the expensive part of the pipeline.

The launch also demonstrates the stealth-preview playbook. Giving away an unbranded model for free generated trillions of tokens of real-world usage and community guessing about its origin, so the official reveal landed with an audience already formed.

Caveats apply. Every figure here comes from a single developer's firsthand account, and the improvised-joke anecdote is one data point, not evidence of a general trait. But if GLM-5.3-Flash holds up at the quality level its early reviewers claim, its pricing puts real pressure on closed-model incumbents.